Add validation support for llm_generation steps

[wyrdbound]

When LLMs are invoked, their outputs are often error prone and ill-formatted. With validation blocks, we will now be able to ask the engine to validate JSON output is parseable and/or abides by a given schema:

Example:

  - id: determine_saving_throw_type
    name: "Determine Saving Throw Type"
    type: llm_generation
    prompt_id: determine_saving_throw_type
    prompt_data:
      context_summary: "{{ inputs.context_summary }}"
      saving_throw_ability: "{{ variables.saving_throw_ability }}"
    validation:
      type: "json_schema"
      schema:
        type: "object"
        properties:
          type:
            type: "string"
            enum: ["basic", "advantage", "disadvantage"]
          reason:
            type: "string"
            minLength: 10
        required: ["type", "reason"]
    actions:
      - set_value:
          path: "variables.saving_throw_type"
          value: "{{ llm_result.type }}"
      - log_event:
          type: "llm_result"
          data: "Saving Throw Type Determined: {{ variables.saving_throw_type }}, Reason: {{ llm_result.reason }}"